X-ray spectra are usually analyzed by fitting the data to a known physical emission model. This is usually done manually with the help of spectral-fitting programs such as XSPEC, which aim to find the global minimum of a known function in a multi-dimensional parameter space using the method of grid-search. Such methods turn inefficient for an increasing number of spectra. On the other hand, machine learning models have recently achieved tremendous success with remarkable predictive capabilities. They may present an alternative solution to analyze X-ray spectra that could overcome the limits of the existing techniques. In this work, we propose Deep Spectra, a neural network composed of a Fully Convolutional Autoencoder, to denoise and disentangle the X-ray source to different components, and Convolutional Neural Networks to predict the parameters of each spectra component. We train the network using a dataset of simulated spectra from the Active Galactive Nuclei using XSPEC. The proposed approach shows a neat increase in performance on simulated data compared to standard spectra fitting routines. We aim to extend the proposed approach to real-world data from the XMM-Newton Catalogue. However, several challenges need to be addressed and they represent interesting future research directions.